Diabetic retinopathy (DR) is a severe complication of diabetes mellitus that can lead to vision loss. This project enhances DR classification using a novel ensemble of deep convolutional neural networks integrating VGG16, ResNet50, and InceptionV3. Techniques such as depthwise separable convolutions, residual learning, and inception modules were incorporated to improve accuracy and efficiency. With a dataset from Kaggle, preprocessing and data augmentation were applied, and results showed notable improvements—particularly after restructuring the classification into binary categories. The final ensemble model, deployed via a PyQt-based GUI, achieved a peak F1 score of 0.946, indicating high sensitivity and specificity in distinguishing DR from non-DR cases.
This project demonstrates the efficacy of CNN-based ensemble learning for medical image classification, especially for diabetic retinopathy. While the initial multi-class models showed moderate performance, restructuring the task into a binary classification significantly enhanced sensitivity and overall accuracy, making the model suitable for early DR screening tools.
The model's performance relies heavily on well-annotated data, and generalizability may be limited when deployed on unseen datasets from varied imaging devices. The ensemble strategy also introduces higher computational complexity.
Future research will explore transfer learning from medical-specific datasets, integration with mobile health platforms, and real-time DR classification embedded in fundus camera hardware. Additionally, federated learning and cross-dataset adaptation will be investigated.